| Intel iMac IDL performance [message #47652] |
Mon, 27 February 2006 13:09  |
K. Bowman
Messages: 330
Registered: May 2000
|
Senior Member |
|
|
Apple loaned us an Intel Dual-Core iMac for a few days for testing. Here is a
quick comparison:
Intel system specs:
2 GHz Intel Core Duo (2 cpus)
2 GB DDR2 SDRAM
667 MHz bus
OS X 10.4.5
PowerPC system specs:
2.5 GHz PowerPC G5 (4 cpus)
2 GB DDR2 SDRAM
1.25 GHz bus
OS X 10.4.5
We installed the Mac (PowerPC) version of IDL on both. The Intel runs IDL via
emulation software (Rosetta).
My IDL benchmark code (dominated by 3-D interpolation, random memory access):
PowerPC 31 s
Intel iMac 61 s
I played with the IDL demo programs on the Intel iMac and everything that I
tried ran fine. Basic interactive IDL performance is very quick.
All in all, IDL seems to run fine. Performance is quite respectable for an
emulated system. Native IDL performance (when available) could be comparable to
the G5.
Ken Bowman
|
|
|
|
|
|
|
|
|
|
| Re: Intel iMac IDL performance [message #47776 is a reply to message #47652] |
Tue, 28 February 2006 12:55  |
JD Smith
Messages: 850
Registered: December 1999
|
Senior Member |
|
|
On Tue, 28 Feb 2006 14:28:26 -0600, Kenneth Bowman wrote:
> In article <pan.2006.02.27.22.35.29.385927@as.arizona.edu>,
> JD Smith <jdsmith@as.arizona.edu> wrote:
>
>> Good news. Can you try running your benchmark a few time, Ken? Rosetta
>> is not an emulator, but a caching code translator. When it encounters
>> code it has already translated, it simply uses its cached version of
>> that, which should run somewhat faster, so it's not unusual to have the
>> second and later runs of a given benchmark speed up. Can you also run:
>>
>> IDL> time_test3
>>
>> a few times? On my PB G4, that takes 3.6s/0.13s total/geom. mean.
>> Sadly, I expect the iBook Intel/MacBook Pro to beat these numbers even
>> under Rosetta. One other good one to try:
>>
>> IDL> a=randomu(sd,100L*!CPU.TPOOL_MIN_ELTS) IDL> t=systime(1) &
>> a=sqrt(a)/(a>0.5) & print,systime(1)-t
>>
>> which shows how well the threading is working on ~40MB of data. On my
>> PBG4, this takes 1.8s.
>>
>> Thanks,
>>
>> JD
>
> Hi, JD.
>
> I ran JD's benchmark, along with time_test3 and my personal benchmark.
> The results are summarized here:
>
> http://idl.tamu.edu/mac_bench.php
>
> I ran all tests 3 times. Variations between individual runs was at the
> 10% level. (Re-running did not produce significant changes in speed.)
>
> The Intel iMac is faster than my (relatively new) PowerBook G4, but slower
> than a high end G5 desktop.
>
> Multi-threading on the quad-processor G5 seems to work quite well.
>
> I ran a few other non-IDL tests. TeX, with the TeXshop front-end, is
> amazingly fast.
Thanks, Ken. As anticipated, it seems the PBG4 is worse at running
IDL PPC code than an emulated Intel Core Duo at 20% higher clock
speed. Ouch. You might add (Rosetta) or something to that Intel iMac
column in case your link turns up on Google for the ever popular
"Intel iMac benchmark" search. Also, can you list the IDL version?
When 6.4 or 6.3.x or whatever comes compiled for Intel, we can re-do
things. My bet: faster than the quad-G5 in time_test3, slower (but
not by much) in my thread-heavy test. Makes me want to find someone
to revive the IDLSPEC of years past. Anyone?
JD
|
|
|
|
| Re: Intel iMac IDL performance [message #47785 is a reply to message #47652] |
Tue, 28 February 2006 09:07 |
JD Smith
Messages: 850
Registered: December 1999
|
Senior Member |
|
|
On Mon, 27 Feb 2006 21:47:15 -0500, Robert Moss wrote:
> JD Smith wrote:
>> On Mon, 27 Feb 2006 15:09:53 -0600, Kenneth Bowman wrote:
>>
>>> Apple loaned us an Intel Dual-Core iMac for a few days for testing. Here is a
>>> quick comparison:
>>>
>>> Intel system specs:
>>> 2 GHz Intel Core Duo (2 cpus)
>>> 2 GB DDR2 SDRAM
>>> 667 MHz bus
>>> OS X 10.4.5
>>>
>>> PowerPC system specs:
>>> 2.5 GHz PowerPC G5 (4 cpus)
>>> 2 GB DDR2 SDRAM
>>> 1.25 GHz bus
>>> OS X 10.4.5
>>>
>>> We installed the Mac (PowerPC) version of IDL on both. The Intel runs IDL via
>>> emulation software (Rosetta).
>>>
>>> My IDL benchmark code (dominated by 3-D interpolation, random memory access):
>>> PowerPC 31 s
>>> Intel iMac 61 s
>>>
>>>
>>> I played with the IDL demo programs on the Intel iMac and everything that I
>>> tried ran fine. Basic interactive IDL performance is very quick.
>>>
>>> All in all, IDL seems to run fine. Performance is quite respectable for an
>>> emulated system. Native IDL performance (when available) could be comparable to
>>> the G5.
>>
>> Good news. Can you try running your benchmark a few time, Ken?
>> Rosetta is not an emulator, but a caching code translator. When it
>> encounters code it has already translated, it simply uses its cached
>> version of that, which should run somewhat faster, so it's not unusual
>> to have the second and later runs of a given benchmark speed up. Can
>> you also run:
>>
>> IDL> time_test3
>>
>> a few times? On my PB G4, that takes 3.6s/0.13s total/geom. mean.
>> Sadly, I expect the iBook Intel/MacBook Pro to beat these numbers even
>> under Rosetta. One other good one to try:
>>
>> IDL> a=randomu(sd,100L*!CPU.TPOOL_MIN_ELTS)
>> IDL> t=systime(1) & a=sqrt(a)/(a>0.5) & print,systime(1)-t
>>
>> which shows how well the threading is working on ~40MB of data. On my
>> PBG4, this takes 1.8s.
>
> Hmm. Maybe your PB is dialed back to save battery power. My Pentium 4m @
> 2.2 GHz and 512 MB RAM gives this:
>
> IDL> a=randomu(sd,100L*!CPU.TPOOL_MIN_ELTS)
> IDL> t=systime(1) & a=sqrt(a)/(a>0.5) & print,systime(1)-t
> 0.62500000
> 1.92300=Total Time, 0.062429919=Geometric mean, 23 tests.
>
> I did run these a couple of times to remove the memory allocation time
> you typically see the first time through. Still, I'm surprised.
Yes, IDL performance on G4's is pretty pathetic. Much better on G5's.
The excuse seems to be gcc, which I believe is used to compile IDL on all
Unix platforms. So really, the advantage for IDL from moving to
PowerPC->Intel will be larger than average, especially for laptop owners.
JD
|
|
|
|
| Re: Intel iMac IDL performance [message #47930 is a reply to message #47652] |
Mon, 13 March 2006 09:59 |
Wolf Schweitzer
Messages: 21
Registered: October 2001
|
Junior Member |
|
|
JD Smith wrote:
> This assumes TPOOL_MIN_ELTS=100000. Setting tpool_min_elts with CPU will
> reset this, which will make the size of the vector much smaller, and make
> this somewhat artificial (though I don't doubt a factor of 10, really). I
> guess I should have put a:
>
> cpu,tpool_min_elts=100000
>
> first, to even the playing field.
>
> JD
I did set the vector to 100000 (which rids me of depending on that
assumption with the threadpool minimal elements - setting being constant).
Then I vary TPOOL_MIN_ELTS until I find the fastest speed. I personally
would see no point in recording an artificially slow speed just because
for a given machine / task, the TPOOL_MIN_ELTS is suboptimal. So you'd
first seek the best speed, and record that.
Below my "tweaked" version.
Regards, Wolf.
---
pro jdstest
cpu,/reset
a=randomu(sd,100L*!CPU.TPOOL_MIN_ELTS)
t=systime(1) & a=sqrt(a)/(a>0.5) & ri =systime(1)-t
print, 'initial result ', ri, ' @ tpool ', !cpu.tpool_min_elts
bs= double(5)
p = 100
pool=0
for n = 0.,30. do begin
cpu,tpool_min_elts=bs^n
p=100
for i = 1,32 do begin
a=randomu(sd,100L *100000) ;*!CPU.TPOOL_MIN_ELTS)
t=systime(1) & a=sqrt(a)/(a>0.5) & r =systime(1)-t
if r lt p then begin
print, ' found new optimum at ',r, ' seconds @ tpool_min_elts ',bs^n
p = r
pool=n
end
end
end
print, ' final results ', p, ' @ tpoolminelts ', bs^pool
print, ' performance gain through tweaking tpool variable : new jd test
runs at percentage of ', p/ri * 100., ' %'
end
|
|
|
|
| Re: Intel iMac IDL performance [message #48087 is a reply to message #47930] |
Mon, 20 March 2006 08:19 |
JD Smith
Messages: 850
Registered: December 1999
|
Senior Member |
|
|
On Mon, 13 Mar 2006 18:59:00 +0100, Wolf Schweitzer wrote:
> JD Smith wrote:
>> This assumes TPOOL_MIN_ELTS=100000. Setting tpool_min_elts with CPU will
>> reset this, which will make the size of the vector much smaller, and make
>> this somewhat artificial (though I don't doubt a factor of 10, really). I
>> guess I should have put a:
>>
>> cpu,tpool_min_elts=100000
>>
>> first, to even the playing field.
>>
>> JD
>
> I did set the vector to 100000 (which rids me of depending on that
> assumption with the threadpool minimal elements - setting being constant).
>
> Then I vary TPOOL_MIN_ELTS until I find the fastest speed. I personally
> would see no point in recording an artificially slow speed just because
> for a given machine / task, the TPOOL_MIN_ELTS is suboptimal. So you'd
> first seek the best speed, and record that.
Because that's somewhat unrealistic, given how the optimum variable will
change depending on the code executing. I think 100000 is a good
conservative choice, though it might be worth trying half that and double
that.
JD
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
